Causal inference and non-randomized experiments

Abstract

Traditionally, machine learning and artificial intelligence focus on problems of diagnosis or prognosis. Answering questions on whether a patient might have a certain disease (diagnosis) or is at risk of future disease (prognosis). In addition to these problems, one might be interested in identifying causal factors which can provide information on how to change disease onset or disease progression. In this chapter we introduce the potential outcomes framework, which provides a structured way of conceptualizing questions on causality. Using this framework we discuss how randomized and non-randomized experiments can be conducted, and analyzed, to obtain estimates of the likely causal effect an exposure may have on an outcome.